Training with several hidden layers

[VeliBaba]

Hi! I have some questions about faster-rnnlm. There it is possible to use several hidden layers during training. My questions are:

1. Which of them is used for recurrent part?
2. Does it use those hidden layers during decoding or computing entropy? Thanks!

[akhti]

Hi!

1. All of them. The output of one layer is the input for the next one. For instance, if you have two tanh layers then the network looks like this: h1_t = tanh(xt + W1 * h1{t - 1}) h2_t = tanh(U * h1t + W2 * h2{t - 1})
2. Yes, it does.

[VeliBaba]

Ok. Is the ouput of the last hidden layer used at the input of the next neural network?

[akhti]

What is 'next neural network'? If you mean next timestamp (next word), then the answer is yes.

[VeliBaba]

Yes, I mean this. Ok, thanks

[VeliBaba]

Is it good at performance using several hidden layers instead of a single hidden layer? Which is better: use a single hidden layer with size 400, or 4 hidden layers with size 100?

[akhti]

First, when you increase layer in 4 times, training/evaluation time (in theory) is increased in 16 times (4 squared). So it's more resonable to compare 1 layer of size 400 with 4 layers of size 200. However, I would recomment to train a shallow network with a single first.

[VeliBaba]

Hi! I have two different toolkits for training of the rnnlm: the first one is rnnlm-hs-0.1b (Ilya-multithreading), and the second one is faster-rnnlm.The faster-rnnlm is faster than rnnlm-hs-0.1b about 3 times with the same options. Is it expectable that valid entropy at the end of training may be worse with faster-rnnlm than rnnlm-hs-0.1b?

[akhti]

It's expected that the entropy will be more or less the same.